tether
Server Configuration
Describes the environment variables required to run the server.
| Name | Required | Description | Default |
|---|---|---|---|
No arguments | |||
Instructions
Guidance the server publishes about itself, which clients place ahead of the tool catalog so the model reads it before choosing anything.
Tether gives you the user's real, logged-in browser tabs. Prefer it to a separate automated browser for anything the user would do in their own browser. You can only use tabs the user shared (and tabs those pages open), unless they approve bypass mode. Start with list_tabs; if nothing is shared, call start_session, which opens your own window in their browser after they approve. Then snapshot to read the page and get element refs. Act with click, type, fill_form; each action returns a fresh snapshot. Use network_requests and console_messages to debug. The user can send you change requests (text and images) from the Tether popup: read them with get_requests. To restyle a site, preview with apply_css, check with screenshot, then save_style; use save_script for behaviour. Before working on a site, site_map shows what Tether recorded about how the user uses it. Ask before submitting payments, sending messages or deleting anything. Call end_session when you are done with your window.
Capabilities
Features and capabilities supported by this server
Protocol revision2025-11-25
| Capability | Details |
|---|---|
| tools | {
"listChanged": true
} |
Tools
Functions exposed to the LLM to take actions
| Name | Description |
|---|---|
| list_tabsA | List the browser tabs you can use. Normally these are the tabs the user shared by clicking Tether, plus tabs you or those pages opened. If the user has turned on bypass mode, every tab is listed. |
| open_tabB | Open a URL in a new tab in the user's browser. The new tab is shared with you. |
| close_tabC | Close a shared tab. |
| navigateB | Go to a URL in a shared tab, or go back, forward or reload. |
| snapshotA | Read the page as an outline of headings, text and interactive elements. Each element you can act on has a ref like [e14]. Take a snapshot before acting and use the refs in click, type and fill_form. |
| read_textA | Get the visible text of the page, or of one element, as plain text. Good for reading articles, tables and results. |
| clickC | Click an element. |
| hoverB | Move the pointer over an element, e.g. to open a menu. |
| typeA | Type into a text field or editable area. Replaces the current value unless append is true. |
| fill_formB | Fill several fields at once. Text fields take a string, checkboxes and radios take true or false, selects take the option label or value. |
| select_optionB | Choose one or more options in a select element, by label or value. |
| press_keyB | Press a key, optionally on an element. Examples: "Enter", "Escape", "Tab", "ArrowDown", "Control+a". |
| scrollB | Scroll an element into view, or scroll the page by a number of pixels. |
| wait_forA | Wait until some text appears or disappears, or for a number of milliseconds. |
| screenshotA | Take a screenshot of the visible part of a shared tab. The tab is brought to the front first. |
| run_jsA | Run JavaScript in the page's own context (page globals are visible) and return the result of the last expression, awaited if it is a promise. Works on sites with a strict CSP. Use for reading data the outline does not show or driving page APIs directly. |
| upload_fileA | Attach local files to a file input, as if the user picked them. |
| save_watcherA | Watch a page and notify the user when something they care about happens ("tell me when the price drops below 2000", "when results are posted"). The browser fetches the page on a schedule with the user's cookies. mode=fetch (default) loads the URL without rendering it, so point it at the JSON API you found with network_requests when there is one; check is a function body that receives response = { status, url, text, json, contentType } and has no DOM. mode=tab opens the page in a background tab for a moment, for pages built by JavaScript; check runs in the page with full DOM access. check either returns a value (the user is notified when it changes from the last run) or { notify: true/false, message, value }. Notifications go to a browser notification and, with notify including "ntfy", to the user's ntfy topic. Test with run_watcher. |
| list_watchersB | List watchers with their last result. |
| run_watcherA | Run a watcher now and return what its check produced. Set notify to true to send the notification if it fires; by default this is a dry run that still records the value as the new baseline. |
| delete_watcherC | Delete a watcher. |
| list_mediaA | List the media on a shared tab: images, video and audio (with size, length, playing state), canvases and CSS background images, plus HLS/DASH streams and media files the page loaded (from the network log). Each has an id like m3 for save_media. Streams only appear after the page has requested them, so start playback first. |
| save_mediaA | Save an item from list_media (or any URL) to disk under ~/.tether/media//, using the browser's cookies. Handles normal files, blob: and data: URLs, canvases, and HLS/DASH streams (through ffmpeg). Returns the file path. |
| video_frameA | Get the current frame of a video on a shared tab as an image. Falls back to a cropped screenshot when the site doesn't allow reading the video. |
| wait_for_downloadA | After clicking something that downloads a file (an invoice, a CSV export, a PDF), wait for it to finish and get its path on disk. Downloads that started in the last 15 seconds count, so call it right after the click. |
| download_fileA | Download a URL through the user's browser, with their cookies, and get the saved file's path. Limited to sites of shared tabs unless bypass mode is on. |
| list_downloadsB | List files downloaded while tabs were shared with you (not the user's whole download history). |
| wait_for_userA | Hand a shared tab to the user and wait for them. With message, the page shows it with a Done button and brings the tab to the front: use it for CAPTCHAs, 2FA codes, logins, payment confirmation or any choice that is theirs. Without message, it waits for the user to resume a tab they paused (actions on a paused tab fail with a note saying so). Returns when they press Done or Resume, or after the timeout. |
| start_sessionA | Start working in the user's own browser (their logins and cookies) when no tab is shared yet. mode=window (default) opens a separate window for you, so the user's tabs are left alone; every tab you open in it is yours. mode=current asks to use the tab the user is looking at, for tasks about the page they're on. The user approves in a prompt (up to two minutes). Returns the tab handle and a snapshot. |
| end_sessionA | Finish a session you started: closes the window start_session opened for you (all of them if you don't name one). With tab, stops using that shared tab instead, leaving it open for the user. |
| request_bypassA | Ask the user for bypass mode, which gives you every open tab in the browser instead of only the shared ones. The browser shows the user a prompt with your reason and waits up to two minutes for an answer. Only ask when the task needs other tabs. |
| network_requestsA | List network requests made by a shared tab since it was shared (newest last): id, method, status, type, time, size, URL. Use network_request with an id to see headers and bodies. |
| network_requestA | Show one captured request in full: request and response headers, request body and response body. |
| console_messagesA | Read console output and uncaught errors from a shared tab since it was shared. |
| cookiesA | List the cookies the browser would send to the tab's current page, including httpOnly ones. |
| list_scriptsA | List the user's userscripts: ones you saved and ones they imported. They live in ~/.tether/scripts as .user.js files, appear in the Tether popup on matching sites, and work unchanged in Tampermonkey or Violentmonkey. |
| save_scriptA | Save a reusable userscript. |
| get_scriptB | Read a saved script, to change it. Returns the editable body for scripts you saved, or the whole file for imported ones. |
| run_scriptC | Run a saved script in a shared tab and return its result. |
| install_userscriptA | Install a userscript (.user.js) or user style (.user.css) from a URL, such as Greasy Fork, OpenUserJS or userstyles.world, or from pasted text. It is saved unchanged. @require'd libraries and style preprocessors (Stylus, Less) are not supported. |
| get_requestsA | Read change requests the user sent from the Tether popup: what they want changed on a site, often with reference images (screenshots, mockups, mood boards). The tab they sent it from is shared with you. Handle open requests, then call resolve_request. |
| resolve_requestA | Mark a change request done and leave a short reply the user sees in the popup (what you changed, and how to turn it off). |
| apply_cssA | Preview CSS on a shared tab without saving it. Each call replaces the previous preview on that tab; navigation clears it. Check the result with screenshot, then save_style. |
| clear_previewC | Remove the CSS preview from a tab. |
| save_styleA | Save CSS that restyles a site on every visit, in every connected browser. It is stored as a UserCSS file (~/.tether/styles/.user.css) that also installs unchanged in Stylus. It shows in the popup on matching pages, where the user can switch it off. Saving an existing name replaces it and bumps its version. Use !important where the site's own rules win. For changes CSS can't do, pair it with save_script (run=auto). |
| list_stylesA | List saved site styles, yours and imported ones. |
| get_styleB | Read a saved style, to change it. Returns the editable CSS for styles you saved, or the whole UserCSS file for imported ones. |
| delete_styleC | Delete a saved style. |
| site_listA | List sites Tether has recorded the user using (clicks, typing, pages, API calls), and sites with saved notes. |
| site_mapA | Summarise everything recorded for a site: its pages, the buttons and links the user actually uses on each (with CSS selectors), the fields they fill, the API endpoints behind it, common paths between pages, and saved notes. Pass page for one page in detail including its last outline. Read this before automating or restyling a site. |
| site_sessionsB | List recorded sessions for a site, or read one session's timeline step by step. |
| site_notesB | Read or write notes about a site that persist for future sessions and other agents: how things work, where things are, gotchas. Without text it reads them. |
| delete_scriptC | Delete a saved script. |
Prompts
Interactive templates invoked by user choice
| Name | Description |
|---|---|
No prompts | |
Resources
Contextual data attached and managed by the client
| Name | Description |
|---|---|
No resources | |
TDQS
Scored across 53 tools
Most tools target a clear browser resource or action, and the descriptions give strong boundaries between page reading (snapshot, read_text, screenshot, run_js) and interaction. A few boundaries are softer: network_requests vs. network_request, site_list vs. site_map vs. site_sessions, and save_script vs. install_userscript can require reading descriptions carefully.
Names are consistently lowercase snake_case and mostly follow recognizable verb_noun or category-prefix patterns. There are minor deviations such as bare verbs (click, type, navigate), noun-first site_list/site_sessions, and noun-only network_requests/cookies, but the conventions remain readable and predictable.
The server covers a very broad browser-automation surface, but 53 tools is far beyond a well-scoped set and imposes a significant selection burden. Many tools earn their place functionally, yet the total count is too high for one coherent toolset.
The surface is unusually complete for browser automation: tabs, interaction, page reading, network/console inspection, sessions, styles, scripts, watchers, media, downloads, and site memory are all covered. Minor gaps remain, such as setting/deleting cookies and explicit dialog handling, though run_js can work around many of them.